Modeling Soil Chemical Properties using Remote Sensing, Artificial Neural Network and Spatial Interpolation Techniques

Abstract

Soil fertility evaluation is a basic and important practice in crop production. However, laboratory analysis is a labor-intensive and costly practice. There is an intensive need for automated statistical and spatial models that can save time and cost. The objective of this study was geospatial variability analysis of soil chemical properties pH, organic matter (OM) , Phosphorus (P), potassium (K), and comparison of different techniques for soil chemical properties prediction using classical, geo-statistical, neural network and remote sensing modeling approaches. Classical statistics, geostatistics, ANN and spatial interpolation (SI) were compared in this review. Multiple Linear Regression (MLR), Ordinary Least Square Regression (OLS), GMDH (Group Method of Data Handling) Neural Networks, Inverse Distance Weighting (IDW), Kriging Interpolation, were used to model and predict soil chemical properties. Landsat-8 and Sentinel-2 models were evaluated for predicting soil chemical properties. Soil OM, and pH were modeled with acceptable accuracy using a variety of modelling techniques. MLR results concluded that soil OM can be better studied using Sentinel-2 while soil pH has fewer chances to be modeled and studied using a single sensor alone, a hybrid approach would be better. Spatial regression results concluded that Sentinel-2 performance was better for soil OM and pH prediction. GMDH ANN results showed, soil OM was better predicted using Sentinel-2, while soil pH was better predicted using Landsat-8. Comparison of spatial interpolation technique concluded that no single technique is best; rasters were evaluated using RMSE. Kriging results were better for Soil OM and pH, while IDW models were better for Soil P and K. Conventional soil testing methods are outdated and must be replaced by automated models.